Automatic modulation recognition based on multi-feature fusion and Sig53 dataset

Automatic modulation recognition (AMR) has become a prominent research focus in wireless communications and radio spectrum management. However, most existing studies are limited by small-scale datasets, insufficient diversity in signal types, and inadequate modeling of channel impairments, making it difficult to thoroughly evaluate model performance in realistic scenarios. To address these limitations, we propose a novel AMR method based on multi-feature fusion and introduce a Multi-Channel Transformer (MCTrans) model that integrates convolutional neural networks with attention mechanisms. Experiments on the large-scale Sig53 dataset demonstrate that fusing IQ, amplitude-phase (AP), and time-frequency features leads to a more comprehensive signal representation. Compared with EfficientNet-B4 and XCiT-Tiny12 models, MCTrans achieves accuracy improvements of 5.26% and 3.83%, respectively. These findings underscore the enhanced feature expressiveness and recognition accuracy of the proposed multi-feature fusion-based AMR approach.